README.md

---
license: gpl-3.0
task_categories:
- tabular-classification
- tabular-regression
- image-to-3d
- depth-estimation
pretty_name: ACE-Rallies
size_categories:
- n<1K
---

**Paper in the making**

---

# ACE-Rallies Dataset

This dataset was created for the Master's thesis *"From Broadcast to 3D: A Deep Learning Approach for Tennis Trajectory and Spin Estimation"* by Alexandra Göppert at the University Augsburg, Chair of Machine Learning and Computer Vision. 

ACE-Rallies serves as an enriched version of the original TrackNet dataset, supplementing the original broadcast tracking data with manual ball spin annotations and 2D human pose estimations for the hitting player. It was specifically designed to be used as a validation and test set (split 33% to 66% respectively) for the 2D-to-3D trajectory uplifting models available in the [tennisuplifting GitHub repository](https://github.com/xSpaceCoder).

### Original TrackNet Data
The foundation of this dataset is the TrackNet dataset, which consists of broadcast tennis videos and 2D ball tracking data. The original dataset was introduced in [this paper](https://arxiv.org/abs/1907.03698) and downloaded from [Kaggle](https://www.kaggle.com/datasets/sofuskonglevoll/tracknet-tennis).

---

## Folder Structure

The repository is organized into 93 sequentially numbered folders, ranging from **`rally_0000`** to **`rally_0092`**. Each folder represents a single, continuous tennis rally (a sequence of consecutive shots starting with a serve or ball toss). 

---

## Data Structures per Rally

Inside each `rally_xxxx` folder, you will find exactly **seven separate `.npy` files** and **one `info.json` file** containing the isolated data for that specific sequence. 

### The JSON File
*   **`info.json`**: This file contains the essential metadata linking this rally back to the original TrackNet data. The fields include the specific game and clip the trajectory was cut from, as well as the exact **start frame number** and **end frame number** relative to the original TrackNet video.

### The NPY Files
The seven numpy arrays store the spatial, temporal, and pose data for the rally:

*   **`Mint.npy`**: A [3 x 3] array representing the reverse-engineered intrinsic camera matrix.
*   **`Mext.npy`**: A [3 x 4] array representing the reverse-engineered extrinsic camera matrix (containing the rotation matrix and translation vector).
*   **`times.npy`**: A 1D array of length [T] (where T is the total number of frames in the rally) containing the exact timestamps for each frame.
*   **`r_img.npy`**: A [T x 3] array containing the 2D ball tracking data per frame. The three columns represent the **`u`** (horizontal) coordinate, the **`v`** (vertical) coordinate, and a **visibility class**. The visibility class is binary, where 0 means the ball is out of frame/occluded, 1 means visible or hard to spot. The visibility class was directly extracke out of the TrackNet Dataset.
*   **`2dPoseEstimation.npy`**: A [17 x 3] array containing the 2D human pose estimation of the hitting player. The rows correspond to the 17 [COCO-WholeBody keypoints](https://arxiv.org/abs/2007.11858), and the columns represent the **`u`** coordinate, **`v`** coordinate, and a model confidence **`score`**. For whole rallies, this pose is captured at the specific frame where the ball leaves the server's hand (or the first frame if the toss isn't visible).
*   **`spin_class_per_shot.npy` & `spin_class_per_frame.npy`**: These files map the initial ball spin of the shots. The classes are categorized as **1 (topspin)**, **2 (backspin)**, and **0 (no spin)**, with 0 typically used for the initial ball toss. `spin_class_per_shot` provides an array of length [S] mapping one spin class to each of the *S* shots in the rally. spin_class_per_frame.npy has the length [T], which assigns each frame of the video a spin class according to the initial spin value for the respective shot.
*   **`new_trajectory_frame_idx.npy`**: An array of length [S] providing the exact frame indices (time steps) at which each new trajectory begins and the corresponding new spin value can be measured.